The idea is explained here, but in French. Still, it might help to look at the picture there to understand the following explanations. We start from a platonic solid. Its symmetry planes define a triangulation of the unit sphere. We pick one of these triangles and consider the three planes intersecting the sphere on its sides. They define a cone with a triangular basis. By a sucession of reflections about the three planes, we can map any point inside this cone. Let's do this. Now we'll perform a sequence of rotations and of translations. If the point gets out of the cone, we use the appropriate combination of reflexions to map it back inside. We iterate the last two steps a certain number of times. Finally, we draw as solid only those points that stayed within a certain distance of the tip of the cone during the whole process.
This yields some very cool imagery, and the patterns obtained often looks strikingly similar to their 2d analogues. Here is a bunch of links: